[xml] xmlTextReader performance question

[Russell Mok <rmok algorithmics com>]

I am switching to the reader API from XPath to improve 
performance. A lot of time is spent on processing a huge 
list of similar elements. The attributes of each element are 
what I am after. I notice a significant performance gain 
when I replace xmlTextReaderGetAttribute to a sequence of

MoveToAttribute
GetConstValue
MoveToElement

probably because there are very few possible values for the 
attributes in all my test documents. I understand it is very 
difficult to answer performance tradeoff questions and that 
could change in the future. I just hope someone can tell me 
a bit more what is happening here. Is there a hash/set for 
these 'const xmlChar *' strings so that allocation and 
deallocation are minimized? Or is it block allocation? If I 
know every attribute value is unique, would anyone recommend 
not using the MoveToAttribute approach? My program is 
multi-threaded and I really want to minimize 
allocation/deallocation.

I have a separate question. Can I save some 
allocation/deallocation if I use xmlReaderForMemory instead 
of xmlReaderForFile? I am thinking of memory mapping the 
whole file.

And thanks so much for libxml2.

Russell